[Bug] server mode, set resolution but not work

[jasonzhang761213]

Git commit

pattern 1:
--diffusion-model C:\StableDifussionModels\ZImage\TwinFlow_Z_Image_Turbo_exp-Q3_K.gguf --vae C:\StableDifussionModels\ZImage\ae.safetensors --llm C:\StableDifussionModels\ZImage\Qwen3-4B-Instruct-2507-Q2_K.gguf --lora-model-dir C:\StableDifussionModels\LCMLora -H 540 -W 960 --sampling-method euler_a --steps 4 --cfg-scale 1 --scheduler discrete --listen-port 2234

result a 512 * 512 image

pattern 2:
--diffusion-model C:\StableDifussionModels\ZImage\TwinFlow_Z_Image_Turbo_exp-Q3_K.gguf --vae C:\StableDifussionModels\ZImage\ae.safetensors --llm C:\StableDifussionModels\ZImage\Qwen3-4B-Instruct-2507-Q2_K.gguf --lora-model-dir C:\StableDifussionModels\LCMLora --height 540 --width 960 --sampling-method euler_a --steps 4 --cfg-scale 1 --scheduler discrete --listen-port 2234

result a 512 * 512 image

pattern 3:
use post .
model = "sd1.5", prompt = "191,A dimly lit birthday party room, round table with a half-cut cake, scattered confetti. the 22-year-old beautiful woman with a right side braid of red hair, light purple eyes, wearing a green long-sleeved knit sweater paired with a light gray short skirt,, standing still, expression neutral, wearing a white shirt. 25-year-old handsome man, stylish short turquoise hair, brown eyes, denim long-sleeve jacket over white t-shirt, blue denim long pants.,, facing her from across the table, one hand raised mid-air, eyes intense. Soft warm lighting, from side, eye-level, upper body, dutch angle;<sd_cpp_extra_args>{"seed":5503420301715773983}</sd_cpp_extra_args>", n = 1, size = "960x540", response_format = "b64_json" }

result a 960 * 544 image （ 544 ， not 540）

Operating System & Version

Windows10

GGML backends

CUDA

Command-line arguments used

--diffusion-model C:\StableDifussionModels\ZImage\TwinFlow_Z_Image_Turbo_exp-Q3_K.gguf --vae C:\StableDifussionModels\ZImage\ae.safetensors --llm C:\StableDifussionModels\ZImage\Qwen3-4B-Instruct-2507-Q2_K.gguf --lora-model-dir C:\StableDifussionModels\LCMLora -H 540 -W 960 --sampling-method euler_a --steps 4 --cfg-scale 1 --scheduler discrete --listen-port 2234

Steps to reproduce

just use command line or web api as above , no one give the correct image output with expected resoltion.

What you expected to happen

generate 960 * 540 image

What actually happened

image resolution not correct

Logs / error messages / stack trace

No response

Additional context / environment details

No response